Landslide susceptibility mapping based on convolutional neural network and conventional machine learning methods

: Landslide susceptibility mapping (LSM) is a useful tool to estimate the probability of landslide 38 occurrence, providing a scientific basis for natural hazards prevention, land use planning, and economic 39 development in landslide-prone areas. To date, a large number of machine learning methods have been 40 applied to LSM, and recently the advanced Convolutional Neural Network (CNN) has been gradually 41 adopted to enhance the prediction accuracy of LSM. The objective of this study is to introduce a CNN 42 based model in LSM and systematically compare its overall performance with the conventional machine 43 learning models of random forest, logistic regression, and support vector machine. Herein, we selected 44 the Jiuzhaigou region in Sichuan Province, China as the study area. A total number of 710 landslides and 45 12 predisposing factors were stacked to form spatial datasets for LSM. The ROC analysis and several 46 statistical metrics, such as accuracy, root mean square error (RMSE), Kappa coefficient, sensitivity, and 47 specificity were used to evaluate the performance of the models in the training and validation datasets. 48 Finally, the trained models were calculated and the landslide susceptibility zones were mapped. Results 49 suggest that both CNN and conventional machine-learning based models have a satisfactory performance 50 (AUC: 85.72% - 90.17%). The CNN based model exhibits excellent good-of-fit and prediction 51 capability, and achieves the highest performance (AUC: 90.17%) but also significantly reduces the salt- 52 of-pepper effect, which indicates its great potential of application to LSM. 53

3 prediction, landslide susceptibility mapping (LSM) is commonly considered as the first stage in the 66 hazard management and can provide a scientific guideline for further disaster management, as it 67 calculates the probability of landslide occurrence and identify the landslide zone of a specific area.

68
Over the past years, various methods have been proposed and applied for dealing with LSM, 69 including qualitative and quantitative methods. Qualitative methods are knowledge-driven based 70 methods that dependent on the experience of a geomorphologist (Lucà, Conforti, and Robustelli 2011; 71 Tien Bui et al. 2016). These methods are gradually being replaced by quantitative methods. The  Reza et al. 2020;Zare et al. 2013). In addition, 82 hybrid models integrated by statistical learning algorithms and machine learning provide more options 83 for LSM. These models include, but are not limited to ANN-fuzzy, fuzzy weight of evidence, EBF-fuzzy 84 logic, Adaptive neuro-fuzzy inference system, and rotation forest (Chen et al. 2018;Zhang et al. 2019).

85
More recently, in addition to the above mentioned methods, the convolutional neural network 86 (CNN) based model has been applied to LSM (Pham et al. 2020;Sameen, Pradhan, and Lee 2020;Wang, 87 Fang, and Hong 2019; Yi et al. 2020). As a powerful deep learning technique, CNN has shown excellent 88 performance in image classification and identification. The attractive capability of CNN is that it can 89 automatically extract robust and general features from convolutional layers and pooling layers without 90 excessive parameters (Zhao, Du, and Emery 2017). As LSM is essentially a binary problem that mines 91 the relationship between landslide predisposing factors and landslides from the dataset and classifies the 92 corresponding pixels, CNN could also be appropriate for LSM. To the best of our knowledge, according 93 to the dimension of data representation, current implementation of CNN based models in the LSM-related 94 literature can be divided into two categories: (1) converting landslide predisposing factors into 1D or 2D environmental information of landslide locations and do not exploit the spatial information, while 3D 97 data contains the information of landslides and surroundings (Yi et al. 2020). A detail comparison study 98 of these techniques can be found in the reference

166
PGA is an important indicator that reflects the relationship between co-seismic landslide density 167 and the earthquake, and also a necessary factor in landslide susceptibility mapping (Meunier, Hovius, 168 and Haines 2007). In this study, the PGA data were adapted from USGS (https://earthquake.usgs.gov/), 169 which range from 0.12 to 0.26 g, being classified into five groups: (<0.12, 0.12-016, 0.16-0.20, 0.20-170 0.24, >0.24) ( Fig.2(a). In LSM, TWI is also a frequently-used factor derived from the digital elevation 171 model (DEM) that quantifies topographic control on hydrological processes (Sørensen, Zinko, and 172 Seibert 2006). As shown in Fig. 2(b), TWI used in this work was divided into five categories, with the 173 values of <5.88, 5.88-7.56, 7.56-10.32, 10.32-15.51, 15.51-24.38. The stability of a slope around a river 174 will be significantly affected by the fluctuation of water in river (Çevik and Topal 2003;Saha, Gupta, 175 and Arora 2002). The roads can affect the spread and size of landslides. Therefore, the distances to rivers 176 and roads derived from a topographical map are considered as predisposing factors (Fig. 2(c)-(d)). NDVI 177 is a crucial factor concerned with the slope stability, especially in mountain areas. Fig. 2(e) shows the 178 NDVI is divided into five categories, with values ranging from -0.50 to 0.88. Land use is another common 179 factor contributing to landslides. Using the supervised classification method, the land use was classified 180 into six categories namely water, construction land, bare land, dense forests, sparse forests, grass land, 181 and others ( Fig. 2(f)). In this study, NDVI and land use were derived from the pre-seismic Sentinel-2 7 data on the Google Earth Engine (GEE). Geological factors play a significant role in landslide 183 susceptibility, including lithology and distance to faults. The original lithology and faults from a 184 geological map were provided by local authorities with 1:500000 scale. The lithology of the study area 185 mainly contains five groups, including group 1: medium thick bedded sandy argillaceous limestone, 186 dolomitic limestone (Carboniferous), group 2: sandy slate, argillaceous limestone and massive quartz 187 sandstone (Devonian), group 3: clastic limestone (Lower Permian), group 4: glacial gravel and proluvium 188 (Quaternary), and group 5: Medium-thick layered limestone dolomite with argillaceous limestone 189 (Triassic) ( Fig. 2(g)). The distance to faults was generated using a Buffer tool in ArcGIS, which was 190 classified into <300 m, 300-600 m, 600-900 m, 900-1200 m, 1200-1500 m, >1500m ( Fig. 2(h)).

191
Morphological factors including elevation, slope angle, slope aspect, and TRI were derived from ASTER

200
Having the landslide inventory and predisposing factors, the next step is to format and integrate 201 these data into datasets for further modeling. As shown in Fig. 3(a), all layers of 12 landslide predisposing 202 factors were stacked together to form a tensor with the size of 12×w×h, where w, h represent the length 203 and width of the entire study area, respectively. Then, by overlaying the landslide inventory with the 204 factor tensor, the specific pixel corresponding to each landslide location was obtained. Note that the grid 205 size of all factor layers and the landslide inventory should be the same to ensure that they can be pixel-206 by-pixel. As mentioned earlier, the grid size of all raster data in this study was 30 × 30 m. However, the 207 factor tensor had different numerical ranges for each dimension. For instance, the slope aspect was dived 208 into 9 groups while elevation divided into 6 groups. Therefore, it was essential to normalize each 209 dimension of the factor tensor for improving the convergence speed and accuracy of the machine learning 210 algorithm (Casella et al., 2013). 8 Fig. 3(b) shows the landslide and non-landslide locations extracted from the factor tensor which 212 were used in the deep learning model. The cell corresponding to each landslide location is taken as the 213 center and then expanded into a raster with a size of n × n. Each cell is assigned a value that contains the 214 data of all factor layers. In this way, more environmental information around the landslides can be 215 considered for further modeling as opposed to just use one grid (Yi et al., 2020). Similarly, the raster of 216 non-landslide location is extracted from the tensor data. The size of raster used for learning should be set 217 according to specific demand (H. Wang et al., 2020). In this study, the size of landslide and non-landslide 218 raster was 17 × 17 using trial and error approach. A total of 1420 12-dimensional training and validation 219 tensors extracted from landslide inventory were generated. Table 1 summarizes the CNN datasets we 220 produced in the study area.

369
The CNN and conventional machine learning models were developed by Pytorch and sklearn, 370 respectively. The hardware environment of this study is that a personal computer with 6GB graphic card 371 GTX1660Ti, a 2.6 GHz Intel(R) Core (TM) i7-9750H CPU, and 16 GB of RAM.

376
The goodness-of-fit of the susceptibility models was evaluated using the training dataset, SRC-377 AUC, and statistical measures. The results are shown in Fig. 5 (a) and ArcGIS10.6 (Fig. 6). The landslide susceptibility zones that indicate the ratio of each susceptibility level 408 to the whole study area were used to qualitatively analyze the landslide susceptibility maps (Fig. 7).

409
In landslide susceptibility map of the CNN model, it was observed that VHS area was relatively 410 concentrated with 17% of the study area, mainly distributed in southeast and northeast, and most 411 landslide points accurately fall into it. The same distribution of VHS area was observed in the models of 412 LR and SVM. Whereas, the VHS and LS area were lower than those of CNN models, and more than 413 50% of the area were calculated as three susceptibility levels ranging from low to high. Results of RF 414 models are nearly similar to LR and SVM but this model is not sensitive to VHS area, which accounts 415 only about 2.5% of the study area.